What we know,
you know too

AI system that understands requests, takes actions (tools, skills) and escalates to humans. At Agenivo: configurable per use case and channel-agnostic.

Regulates data processing by a service provider on behalf of the controller. Agenivo signs a full DPA per Art. 28 GDPR with every customer.

Cloud Computing Compliance Controls Catalogue from the German BSI. Standard for secure cloud usage in Germany.

Encryption model where the customer keeps the master keys in their own HSM. The provider cannot decrypt data.

Clarifying Lawful Overseas Use of Data Act. Allows US authorities access to data processed by US companies – even when servers sit outside the US.

Time a container or pod needs after start to answer the first request. Eliminated at Agenivo Strato via warm pool (< 80 ms).

Numeric value (0–1) for a model's certainty about its answer. At Agenivo Logica: configurable threshold for human escalation.

Numeric vector that captures the semantic meaning of a text. Basis for retrieval, similarity search and RAG.

Regulation (EU) 2024/1689. Risk-based regulatory framework for AI systems in the EU. High-risk systems (Annex III) face strict duties.

Mechanism by which an AI agent passes a conversation to a human – on uncertainty, regulatory obligation or customer request.

Switch to a standby region or instance on failure. At Agenivo Strato: multi-region failover in under 30 seconds.

Measures how strictly a generated answer sticks to retrieved sources. Low faithfulness = hallucination.

Training a base model further on domain-specific data. Agenivo prefers RAG over fine-tuning – faster, cheaper, more reversible.

Regulation (EU) 2016/679. Governs protection of personal data in the EU. Agenivo meets it architecturally via privacy-by-default and data minimisation.

LLM output that sounds plausible but is factually wrong. Reduced by RAG, hybrid reasoning and confidence routing.

Hardware-based key module for cryptographic operations. Keys never leave the module. Prerequisite for BYOK.

Combination of a rule-based symbol engine and LLM inference. At Agenivo Logica: reduces hallucinations in regulated domains by 64%.

A model's prediction on an input. For LLMs: token-by-token generation. Latency and cost are measured in tokens.

International standard for information security management systems (ISMS). Agenivo certified since 2024.

Large language model. Statistical model that understands and generates language. Agenivo supports OpenAI, Anthropic, Mistral, Google, xAI and EU providers.

A single platform instance serves multiple tenants (customers, business units) with full data isolation. At Agenivo: row-level security and tenant-specific encryption.

Documentation of a model: training data origin, risk class, accuracy metrics, known limitations. Mandatory under EU AI Act for high-risk systems.

Technique where an LLM retrieves relevant documents from a knowledge base before answering. Cuts hallucinations, enables source citations.

Database mechanism that filters access to table rows per tenant. Prerequisite for safe multi-tenancy.

Maximum tolerated data loss in time. At Agenivo Strato: < 5 minutes, usually 0 seconds (synchronous replication).

Maximum tolerated recovery time after an outage. At Agenivo Strato: < 30 seconds multi-region failover.

Defines what an agent can do: booking, lead capture, knowledge lookup, escalation. At Agenivo, configurable per agent.

Cloud offering hardened against CLOUD Act access and operating under EU law. Examples: T-Systems Sovereign Cloud, IONOS, OVHcloud.

Industry standard for information security in the automotive sector. Prerequisite for supplier relationships with OEMs.

Sub-word unit in which LLMs process text and inference cost is billed. A German text ≈ 1.3 tokens per word.

Independent authority that issues cryptographic timestamps. At Agenivo Custos: daily anchoring of the audit hash chain in an independent TSA.

Specialised database for high-dimensional vectors with similarity search. Basis for RAG. At Agenivo: pgvector with tenant-specific encryption.

Containers waiting in RAM with the model loaded. Eliminates cold starts. At Agenivo Strato: predictive pre-scaling fills the warm pool.

Storage mode where data cannot be modified after writing. Prerequisite for audit-grade logs.